lora-scripts-next

SD-Trainer — One-click LoRA training GUI for SD / SDXL / Flux / Anima
_{Powered by kohya-ss/sd-scripts, with the familiar Akegarasu GUI experience.}

Download  ·  中文文档  ·  Credits & License

Download SD-Trainer-v2.3.0.7z (~55 MB, includes embedded Python) from Releases, extract, and double-click run_gui.bat.

First launch auto-installs PyTorch + CUDA + all dependencies (~3 GB download). Chinese users get mirror acceleration automatically.

Requirements: Windows 10/11 64-bit, NVIDIA GPU (RTX 20+), ~7 GB disk space.

The Windows portable package (SD-Trainer-v*.7z) does not install Flash Attention 2; training uses xformers or PyTorch SDPA. This is intentional, not a failed install.

For Flash Attention 2, use install from source and follow Flash Attention 2 (source / venv). Portable flash-attn support may come later when embed Python + triton is reliable.

git clone https://github.com/wochenlong/lora-scripts-next.git
cd lora-scripts-next

The browser auto-opens http://127.0.0.1:28000 on launch.

Python version: 3.10 recommended (full compatibility). 3.11–3.12 mostly works. 3.13+ is not supported.

By default the system default browser opens. Use --browser to pick one:

python gui.py --browser chrome
python gui.py --browser edge

Portable users: see the section above — do not pip install flash-attn into python_embeded.

This section is for git clone + venv (or a full Python under python\), with PyTorch 2.7.0 + CUDA 12.8 installed.

Priority for Anima: flash → xformers → torch (PyTorch SDPA).

• Python 3.10 recommended (3.11–3.12 if a matching prebuilt wheel exists)
• 64-bit venv — not the portable python_embeded
• Matching PyTorch: torch==2.7.0+cu128, torchvision==0.22.0+cu128
• Windows: install both triton-windows and flash-attn (flash-attn imports Triton kernels at runtime)

1. Clone the repo and run run_gui.bat on first launch (install-cn.ps1 or install.ps1 creates venv, deps, and tries the flash-attn wheel).
2. On every launch, run_gui.ps1 checks triton + flash_attn; if missing, it installs triton-windows then the prebuilt wheel (failure is non-fatal — falls back to xformers / SDPA).

China mirror first-time install:

powershell -ExecutionPolicy Bypass -File .\install-cn.ps1

International:

powershell -ExecutionPolicy Bypass -File .\install.ps1

Inside an activated venv:

.\venv\Scripts\activate

# 1. PyTorch (if not already installed)
pip install torch==2.7.0+cu128 torchvision==0.22.0+cu128 --index-url https://download.pytorch.org/whl/cu128

# 2. Triton (required on Windows, before flash-attn)
pip install "triton-windows<3.4"

# 3. Flash Attention 2 prebuilt wheel (Python 3.10 example)
pip install https://huggingface.co/lldacing/flash-attention-windows-wheel/resolve/main/flash_attn-2.7.4.post1%2Bcu128torch2.7.0cxx11abiFALSE-cp310-cp310-win_amd64.whl

Use cp311 / cp312 in the wheel filename if that is your Python version.

bash install.bash    # venv + torch/xformers/requirements + optional flash-attn build
bash run_gui.sh

Building flash-attn from source needs a CUDA toolkit and C++ compiler; on failure, xformers / SDPA is used.

python -c "import triton; import flash_attn; from flash_attn.ops.triton.rotary import apply_rotary; print('Flash Attention 2 OK')"

Then run python gui.py and start Anima LoRA training — logs should show attn_mode flash.

• Multi-model — SD 1.5 / SDXL / Flux / Anima all work out of the box
• Anima LoRA training — One-click sidebar entry, supports LoRA / LoKr (LyCORIS) / T-LoRA
• Attention backends — Source/venv: Flash Attention 2 when available (Windows prebuilt wheels). Portable package: xformers / PyTorch SDPA only (flash-attn not supported yet)
• T-LoRA — Timestep-Dependent LoRA with dynamic rank and orthogonal init (paper)
• Train Monitor — Auto-opens with GUI, TensorBoard-backed Loss / LR scalar cards, key training parameter checks, real-time progress, terminal log echo, and preview samples
• Built-in TensorBoard — Accessible from the sidebar, no extra setup
• GPU detection — Detects NVIDIA / AMD GPUs on first run; AMD users get a friendly notice with ROCm guidance
• AutoDL ready — Dedicated startup script start_autodl.sh

_{TensorBoard-backed Loss / LR scalar cards in the 6008 Train Monitor}

_{Preview samples update directly in the monitor page}

_{Training logs are shown in both CMD and the monitor page}

See CONTRIBUTORS.md for the full list of contributors and upstream credits.

_{Maintainer: @wochenlong}